This section is the video's sponsored segment — Neon is explicitly named as the sponsor (8:30). The underlying architectural point (one database for vectors + relational) stands on its own; treat the vendor specifics accordingly.

The setup

  • Neon = serverless Postgres — full relational database power, plus pgvector support, so embeddings live alongside relational data (8:30).
  • Payoff: one system instead of two (one for vectors, one for metadata) (8:30).

What that unlocks

  • Semantic search and date filtering in the same query (8:54).
  • Join chunks back to their parent documents (8:54).
  • Track versions and timestamps (8:54).

Serverless scaling

  • No paying for idle capacity when the RAG system isn't in use; no scrambling to provision when traffic spikes (8:54, 9:06).
  • Scales down to zero with no traffic, spins back up in milliseconds on a request (9:06).

The Databricks data point

  • Neon was acquired by Databricks for about $1 billion (9:06).
  • Databricks' stated reason: AI agents. Per the video, they said over 80% of databases created on Neon are now provisioned by AI agents, not humans (9:23).
  • Framing: serverless + instant provisioning is a natural fit for agentic workflows (9:42).

Database branching — the RAG-specific win

  • Postgres underneath means you can branch the database: instantly copy your entire knowledge base to test a new chunking strategy or a new embedding model without touching production, then merge back when satisfied (9:42).
  • Described as "Git for your database" (10:07).